Tanpo Retention
A compact customer-retention specialist (~1.2B) for onboarding, activation, churn-save, customer-success playbooks, health-score follow-up, renewals, expansion, escalation, identity guardrails, and winback workflows — built for local and inexpensive deployment.
Creator: d4rkninja
Collection: Tanpo — Domain Specialists
Original upstream: LiquidAI/LFM2.5-1.2B-Instruct (~1.17B parameters, 32,768-token context, designed for edge/on-device deployment).
Fine-tuning: Unsloth-compatible loading of that checkpoint via hub id unsloth/LFM2.5-1.2B-Instruct (LoRA / PEFT).
This repository hosts the merged Transformers weights (LoRA merged into the base).
Overview
Tanpo is a family of compact domain-specialized business models for local / edge / inexpensive deployment. Different specialists cover different workflows. Tanpo Retention is one specialist in that family (not a frontier or general-purpose model).
Related artifacts:
- LoRA adapter: d4rkninja/tanpo-retention-LoRA
- GGUF quants: d4rkninja/tanpo-retention-GGUF — prefer
Q4_K_Mwhen available
Best For
- Customer-success playbooks, renewal preparation, and risk follow-up
- Onboarding and activation plans tied to time-to-value
- Churn-save, winback, expansion, and escalation drafts
- Health-score interpretation prompts and practical account action plans
Not Designed For
- Fully automated account cancellation, pricing, or customer eligibility decisions
- Fabricating customer facts, impersonation, or unauthorized access to customer systems
- Guarantees of retention, renewal, expansion, or customer outcomes
- General coding or non-retention chat
Why a Specialist Model?
Retention work rewards repeatable structure (signals, diagnosis, intervention, owner, timing, and next step). Specializing a small model for those workflows enables private, low-cost inference without a large general model.
Evaluation
Internal automated domain evaluation (DarkLab harness). Treat as directional, not an industry benchmark.
| Model | Rubric overall |
|---|---|
Base LFM2.5-1.2B-Instruct | 94.2% |
| tanpo-retention | 96.0% |
| Delta | +1.8 percentage points |
Largest gains were on escalation, renewals, and winback at 100%; identity_guardrails is +34.3 percentage points vs base.
Trailed on cs_playbooks (−22.2 percentage points) and health_scores (−5.5 percentage points). Overall still ahead of base; these categories are disclosed.
Artifacts: evaluation/ — COMPARE_RETENTION.md.
Methodology: DarkLab automated domain evaluation. Same prompts and generation config for base vs fine-tune. Not an industry benchmark.
Limitations of this eval: Automated rubrics can reward structure over real-world quality; sample size is small; results may not transfer outside the task distribution.
Example Prompts
- User: Build a 30-day onboarding and activation plan for a B2B SaaS account with a stalled implementation and three user roles.
- User: Draft a renewal-risk action plan from these account signals, separating facts, hypotheses, owner, timing, and next customer conversation.
- User: Create a respectful winback sequence for a customer who churned after an unresolved support escalation; do not invent account facts.
Inference
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "d4rkninja/tanpo-retention"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo,
torch_dtype="auto",
device_map="auto"
)
messages = [
{"role": "system", "content": "You are Tanpo Retention, a practical customer-retention assistant."},
{"role": "user", "content": "Draft a renewal-risk action plan from these account signals, separating facts, hypotheses, owner, timing, and next customer conversation."},
]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Training
Verified from published adapter configs / training artifacts (no unverified hyperparams):
| Field | Value |
|---|---|
| Method | LoRA (PEFT) via Unsloth FastLanguageModel on hub id unsloth/LFM2.5-1.2B-Instruct (Unsloth-compatible of LiquidAI/LFM2.5-1.2B-Instruct) |
LoRA rank (r) | 16 |
| LoRA alpha | 16 |
| LoRA dropout | 0 |
| Bias | none |
| Target modules | Unsloth/PEFT regex targeting attention and MLP projection modules (see adapter_config.json) |
| Task type | CAUSAL_LM |
Merged via PEFT merge_and_unload into full weights in this repo.
Dataset
- d4rkninja/tanpo-retention-sft — format-fixed chat SFT examples (documented on the dataset card).
Limitations
- Specialized: quality drops outside the retention and customer-success workflow distribution.
- ~1.2B scale: limited world knowledge and long-horizon reasoning vs larger models.
- Customer, health-score, renewal, and escalation recommendations can be wrong or incomplete; verify against source systems and account context.
- Eval gains are rubric-based and directional only;
cs_playbooksandhealth_scoresstill trail the base.
Responsible Use
Not a substitute for customer-success judgment or account-system data. Humans must review customer communications, retention offers, escalation plans, and any consequential account decisions. Do not use for deception, harassment, unauthorized access, or discriminatory treatment.
License
license: other / license_name: lfm-1.0
Tanpo merged and GGUF weights are derivatives of LiquidAI/LFM2.5-1.2B-Instruct under the LFM Open License v1.0 (including the commercial Threshold of approximately $10M annual revenue). See the base model card and its LICENSE file. Credit: LiquidAI. Do not treat this stack as Apache-2.0.
Tanpo Family
Tanpo is a family of compact domain-specialized models for focused business workflows.
This specialist is available as:
- Merged:
d4rkninja/tanpo-retention - LoRA:
d4rkninja/tanpo-retention-LoRA - GGUF:
d4rkninja/tanpo-retention-GGUF
Browse all Tanpo specialists: Tanpo — Domain Specialists